FTX's Dead Miners Are Not the Asset — The Interconnect Is

The SHA-256 hash board inside an FTX-era mining container is worth less than the aluminum casing around it. That is not a hot take; it is a hardware constraint. An application-specific integrated circuit built to do exactly one thing — grind SHA-256 — cannot execute a single floating-point matrix multiply. So when the wires carry the headline "FTX's old Bitcoin mining infrastructure is being given a second life powering AI," the machine most readers picture is the wrong machine. What actually changed hands is not compute. It is a substation, a stack of power purchase agreements, a cooling loop, forty acres of dirt with a grid interconnect, and the legal right to pull megawatts off a transmission line during a demand spike. Nobody is recycling ASICs. They are recycling watts. The ledger remembers what the ego forgets: silicon depreciates on a schedule, but capacity appreciates on a curve. I have audited enough of these shells to know the difference. In 2017 I sat in Remix and traced integer overflow bugs through three ERC-20 contracts before launch, and the lesson that stuck was not "code is dangerous." It was that value is never where the marketing points. Value sits in the part of the system that is hardest to replicate. For a Bitcoin mine, that part has never been the miner. Understand the backdrop. FTX collapsed in November 2022, and its estate — overseen by a court-appointed restructuring team in Delaware — has spent the years since liquidating everything that could be liquidated: venture stakes, real estate, the exchange's own token, which is now functionally zero. Mining infrastructure is one line item in a very long schedule of assets. When a headline surfaces that this infrastructure is being repurposed for AI, the real news is not the technology. The real news is that a bankruptcy estate found a buyer, or a lessee, for a physical asset the market had already written down to salvage value. That matters because the mine-to-AI pivot is not a metaphor. It is the dominant re-rating story of the current cycle. Multiple North American miners have signed long-term hosting agreements with AI cloud providers, converting sites built for hashing into sites running high-performance compute. This is external context, not a disclosed fact of the FTX transaction — and the gap between those two categories is exactly where people lose money. Here is the technical core. A modern air-cooled Bitcoin mine and a modern AI data hall share a surprising amount of physical DNA. Both want cheap, firm, high-voltage power. Both want large contiguous land near transmission. Both want redundant fiber. Both need serious cooling, though this is where they diverge hard: a SHA-256 mine can run hot and loud, while a GPU cluster running dense training loads wants liquid cooling and tight thermal control. Retrofitting a mine for that is not a flip. It is a capital project with a GPU procurement tail attached. The reusable layer is what I call the grid stack: interconnection rights, transformer capacity, switchgear, permits, water rights. Interconnection rights are the scarcest asset in the entire chain. In the largest US grid regions, the queue to connect a new large load runs into years, and capacity is rationed. A site that already holds an approved interconnect is, functionally, a monopoly on a number. That is the thing AI demand cannot conjure quickly, and that is the thing FTX's estate happened to be sitting on. The second reusable layer is power economics. A mine lives or dies on its cost per kilowatt-hour. Many were built in power-rich, fiber-poor places — West Texas, the Pacific Northwest, pockets of hydro. Cheap electricity was the whole point. But AI workloads are latency-sensitive in a way Bitcoin mining never was. A hashing rig does not care whether the round trip to its pool is five milliseconds or fifty. A distributed training job, a real-time inference endpoint, a synchronization-heavy cluster — these care a great deal. So the moment you repurpose a remote mine, you inherit its geography, and geography is destiny. The cheapest power is usually the farthest from the fiber. I have watched this shape play out with real money. In 2020 I ran leveraged yield farming on Aave and learned to treat exit liquidity as a first-class variable, not an afterthought. In 2021 I burned two thousand dollars of gas to save fifteen thousand in slippage on a floor sweep, and the takeaway was that friction is a cost you either pay deliberately or pay accidentally. The mine-to-AI trade has the same anatomy. The friction is not the GPUs. The friction is the interconnect queue, the cooling retrofit, and the client contract. Which brings us to what the market is actually pricing. The headline treats "FTX infrastructure powers AI" as a bullish signal for the AI-crypto complex. I read it differently. The estate is a forced seller. Forced sellers do not get fair value; they get the value that clears the docket. If a buyer acquired this asset at a distressed basis and locked a take-or-pay hosting contract with a creditworthy AI tenant, that buyer captured the spread between bankruptcy pricing and market pricing. The FTX creditors captured cash. The AI narrative captured a headline. Three different parties, three different P&Ls. Alpha hides in the friction of chaos — and a bankruptcy estate is the purest chaos there is. Retail sees the ticker and the theme. The desk sees the basis. Now the contrarian read, and it is uncomfortable. The consensus story is "miners become AI landlords, therefore mining assets are underpriced." The blind spot is on the demand side. AI data center buildout across 2024 and 2025 has been funded by a capex cycle that assumes near-linear growth in compute demand. If hyperscaler capex guidance decelerates — and it will decelerate at some point, because every capex cycle does — then the marginal megawatt of converted mining capacity loses its bid. The asset does not become worthless; it becomes a stranded power contract looking for a tenant that stopped calling. Client concentration compounds the problem. AI hosting revenue tends to lean on a handful of Neocloud or hyperscale tenants, so a single non-renewal can gut utilization overnight — the mine's old problem of pool dependency, reborn in a new coat. That is the hidden variable the headline never mentions. The second blind spot is the FTX label itself. Reputation is a balance-sheet item. A site carrying bankruptcy residue — legacy contracts, unresolved litigation, a name that compliance departments flag — trades at a discount. That discount is a tax the buyer pays in diligence and brand repair. Code does not lie, but it does obfuscate, and so does a paper trail that runs through a Delaware courtroom. So what do I actually watch? Not the announcement. The docket. If the facility genuinely changed hands, the sale or lease shows up in public bankruptcy filings, and the counterparty shows up with it. I watch hashrate growth as the tell: if capital is genuinely migrating from hashing to hosting, network hashrate growth should flatten even as mining equities rally — a divergence between the equity story and the physical reality. I watch power purchase agreement prices in the relevant grid region; rising PPA costs erode the very margin the retrofit is meant to capture. I watch hyperscaler capex guidance as the demand-side clock. And I watch the GPU supply chain, because a cooling retrofit is worthless if you cannot source the accelerators. Silence in the order book is louder than noise. Right now the order book on this specific asset is silent — no confirmed counterparty, no disclosed megawatts, no signed tenant. That silence is the signal. It tells me the trade is a thesis, not a position. The real question is not whether old Bitcoin mines can run AI. They can; the physics is settled. The question is whether the AI demand that justifies the conversion survives the next capex air pocket. When the headlines stop being about ruins reborn and start being about utilization rates, we will know which side of the trade the smart money was on.